PoreSkel: Skeletonization of grayscale micro-CT images of porous media using deep learning techniques

Abstract

Skeletonization, a crucial step in pore network modeling, traditionally involves the extraction of skeleton pixels from binarized, segmented X-ray images of porous materials. However, this conventional approach often suffers from user bias during segmentation, potentially leading to the loss of essential image details. This study addresses this limitation by developing deep learning model, called PoreSkel, designed to directly perform skeletonization and distance map extraction from unprocessed grayscale images, thus eliminating the need for additional image processing steps. The model was trained, validated, and tested using an expansive databank of micro-CT images from 20 distinct sandstones, carbonates, and sand pack samples, a total of 10,240 images, with each sample represented by a cube of size 5123. A fifth of these images, specifically 15.6 % from sixteen sandstone and sand pack samples, were used for training, while the remainder served for model validation (4.4 %) and extensive testing (80 %). PoreSkel showed an excellent performance, achieving a mean f1-score of 0.964 for skeletonization and an RMSE of 0.057 for distance map extraction during the testing phase. Our assessments revealed that the model is robust to bias toward the majority class, namely the background pixels. Furthermore, the model showed high generality, maintaining its performance when tested using unseen images from three carbonates and an additional sandstone. Notably, PoreSkel effectively handles disruptions caused often by the presence of minerals in pore spaces and perturbations on pore boundaries - a common challenge for the medial axis technique - resulting in fewer nodes (i.e., pore junctions) and pore coordination numbers, but a higher number of connected skeletons. Therefore, PoreSkel provided a more precise and representative pore structures of porous material that is needed for accurate pore network generation and modeling.